Reections for quantum query algorithms

Ben W. Reichardt
Abstract
We show that any boolean function can be evaluated optimally by a quantum query algo-
rithm that alternates a certain xed, input-independent reection with a second reection that
coherently queries the input string. Originally introduced for solving the unstructured search
problem, this two-reections structure is therefore a universal feature of quantum algorithms.
Our proof goes via the general adversary bound, a semi-denite program (SDP) that lower-
bounds the quantum query complexity of a function. By a quantum algorithm for evaluating
span programs, this lower bound is known to be tight up to a sub-logarithmic factor. The extra
factor comes from converting a continuous-time query algorithm into a discrete-query algorithm.
We give a direct and simplied quantum algorithm based on the dual SDP, with a bounded-error
query complexity that matches the general adversary bound.
Therefore, the general adversary lower bound is tight; it is in fact an SDP for quantum query
complexity. This implies that the quantum query complexity of the composition f (g, . . . , g) of
two boolean functions f and g matches the product of the query complexities of f and g, without
a logarithmic factor for error reduction. It further shows that span programs are equivalent to
quantum query algorithms.
1 Introduction
The query complexity, or decision-tree complexity, of a function measures the number of input bits
that must be read in order to evaluate the function. Computation between queries is not counted.
Quantum algorithms can run in superposition, and the quantum query complexity therefore allows
coherent access to the input string. Quantum query complexity with bounded error lies below
classical randomized query complexity, sometimes with a large gap [BV97, Sim97, Sho97, Aar09],
but for total functions [BBC
+
01] or partial functions satisfying certain symmetries [AA09] the two
measures are polynomially related; see the survey [BW02].
Although the query complexity of a function can fall well below its time complexity, studying
query complexity has historically given insight into the power of quantum computers. For example,
the quantum part of Shors algorithms for integer factorization and discrete logarithm is a quantum
query algorithm for period nding [Sho97]. Unlike for time complexity, there are also strong
information-theoretic methods for placing lower bounds on quantum query complexity. These
lower-bound techniques can be broadly classied as using either the polynomial method [BBC
+
01]
or the adversary method [Amb02,

SS06]. Hyer and

Spalek [H

S05] have surveyed the development

of these two techniques and their multitude of applications. For now, suce it to say that the two
techniques are incomparable. In particular, for the n-input collision problem, the best adversary
lower bound is of O(1), whereas the correct complexity, determined by the polynomial method [AS04]
and a matching algorithm [BHT98] is (n
1/3
).
1
a
r
X
i
v
:
1
0
0
5
.
1
6
0
1
v
1


[
q
u
a
n
t
-
p
h
]


1
0

M
a
y

2
0
1
0
However, Hyer, Lee and

Spalek [HL

S07] discovered a strict generalization of the adversary

bound that remains a lower bound on quantum query complexity:
Denition 1.1 (Adversary bounds). For nite sets C and E, and T C
n
, let f : T E.
An adversary matrix for f is a [T[ [T[ real, symmetric matrix that satises x[[y = 0 for
all x, y T with f(x) = f(y). Dene the adversary and general adversary bounds for f by
Adv(f) = max
0
||
max
j[n]
|
j
|
(1.1)
Adv

(f) = max

||
max
j[n]
|
j
|
. (1.2)
Both maximizations are over adversary matrices , required to be entry-wise nonnegative in Adv(f).

j
denotes the entry-wise matrix product between and
j
=

x,yD:x
j
=y
j
[xy[.
Although the denitions of the two adversary bounds are very similar, the general adversary
bound is much more powerful. In fact, the general adversary lower bound is always nearly tight:
Theorem 1.2 ([Rei09a]). For any function f : T E, with T C
n
, the quantum query complexity
Q(f) satises
Q(f) = O
_
Adv

(f)
log Adv

(f)
log log Adv

(f)
log [C[ log [E[
_
. (1.3)
This surprising upper bound follows from a connection between quantum query algorithms
and the span program computational model [KW93] rst observed in [R

S08] and signicantly

strengthened in [Rei09a, Rei09b]. Note that the original statement [Rei09a, Theorem 10.2] of
Theorem 1.2 restricted to the case [C[ = 2 and included an additional factor of log log [E[this
factor can be removed by [BNRW05, Corollary 3]. Lee has shown that the general adversary bound
of a function with boolean output is stable under encoding the input into binary [Lee09], allowing
the restriction [C[ = 2 to be removed at a logarithmic cost.
Theorem 1.2 and the connection between span programs and quantum query algorithms behind
its proof have corollaries including a query-optimal and nearly time-optimal quantum algorithm
for evaluating a large class of read-once formulas over any nite gate set [Rei09c]. However, is
Theorem 1.2 optimal? The factors of log [C[ and log [E[ are natural, but the log over log log term
is not. It comes from converting a certain continuous-time query algorithm into a discrete-query
algorithm following [CGM
+
09]. This conversion also somewhat obscures the algorithms structure.
It was conjectured that the unnatural log over log log factor could be removed [Rei09a, Conjec-
ture 11.1]. In this article, we conrm the conjecture:
Theorem 1.3. For any function f : T 0, 1, with T 0, 1
n
, the general adversary bound
characterizes quantum query complexity:
Q(f) =
_
Adv

(f)
_
. (1.4)
Theorem 1.3 suces to simplify Eq. (1.3) following the proof of [Rei09a, Theorem 10.2]:
Corollary 1.4. For nite sets C and E, T C
n
, and any function f : T E,
Q(f) =
_
Adv

(f)
_
and Q(f) = O
_
Adv

(f) log [C[ log [E[

_
. (1.5)
2
Theorem 1.3 also allows for obtaining optimal results for the query complexity of composed func-
tions. For f : 0, 1
n
0, 1 and g : 0, 1
m
0, 1, let f g be the function 0, 1
nm
0, 1
dened by (f g)(x) = f
_
g(x
1
, . . . , x
m
), . . . , g(x
(n1)m+1
, . . . , x
nm
)
_
. A bounded-error algorithm
for evaluating f g can be built from composing bounded-error algorithms for f and g; thus,
Q(f g) = O
_
Q(f)Q(g) log Q(f)
_
. However, the logarithmic factor for error reduction can be
removed, and there is a matching lower bound:
Theorem 1.5. Let f : 0, 1
n
0, 1 and g : 0, 1
m
0, 1. Then
Q(f g) =
_
Q(f)Q(g)
_
. (1.6)
Proof. The general adversary bound composes as Adv

(f g) = Adv

(f)Adv

(g) for boolean

functions f and g [HL

S07, Rei09a], so the claim follows from Eq. (1.4).

The algorithm behind Theorem 1.3 is substantially simpler than the algorithm for Theorem 1.2,
although its analysis requires slightly more work. On input x 0, 1
n
, the algorithm consists
of alternating applications of the input oracle O
x
a unitary that maps [i, b to [i, x
i
b, for
i = 1, . . . , n and b 0, 1and a certain xed reection. The reection is about the eigenvalue-zero
subspace of the adjacency matrix A
G
for a graph G derived from a dual SDP for Adv

. This structure
is based on Ambainiss AND-OR formula-evaluation algorithm [Amb07]. A previous algorithm in
Szegedys quantum walk model [Sze04] ran in O(Adv

(f)| abs(A
G
)|) queries, where | abs(A
G
)| is
the operator norm of the entry-wise absolute value of A
G
[Rei09a, Prop. 9.4]. Ambainiss approach
eciently removes the dependence on higher-energy portions of the adjacency matrix A
G
. The
analysis of the algorithm needs to transfer an eective spectral gap for the adjacency matrix of a
related graph into an eective spectral gap for the product of O
x
and the xed reection.
In fact, the input oracle is itself a reection, O
2
x
= 1. Therefore the algorithm consists of
alternating two xed reections, much like Grovers search algorithm [Gro96]. It follows that
every boolean function can be evaluated, with bounded error, optimally in this way. While known
algorithms can in principle be converted into this form [Rei09a, Theorems 3.1, 5.2], we do not know
an explicit closed form for the second reection, e.g., for the collision problem.
The bipartite graph G can be thought of as a span program [KW93], and was constructed
in the span program framework of [Rei09a]. Thus our algorithm is naturally seen as a quantum
algorithm for evaluating span programs. Since the best span program for a function has witness
size exactly equal to the general adversary bound [Rei09a, Rei10], Theorem 1.3 also implies that
quantum computers, measured by query complexity, and span programs, measured by witness size,
are equivalent computational models for evaluating boolean functions. For simplicity, though, we
will not detail this connection further. We will require from [Rei09a] only Theorem 3.2 below, which
without reference to span programs gives an eective spectral gap for a bipartite graph.
Barnum, Saks and Szegedy [BSS03] have given a family of SDPs that characterize quantum
query complexity according to their feasibility or infeasibility, instead of according to the optimum
value of a single SDP. The BSS SDPs work for any specied error rate, including zero. The general
adversary bound is a polynomially smaller SDP, but of course the truth table of a function is
typically exponentially long. Whereas our algorithm uses a workspace of n + O(log n) qubits to
evaluate an n-bit boolean function (by [Rei09a, Lemma 6.6]), n+1 qubits suce by [BSS03]. To the
authors knowledge, neither the BSS SDPs nor Adv

have ever been solved directly for a nontrivial,

asymptotically large family of functions, with better than a constant-factor improvement over the
adversary bound (see [CL08]). However, the easy composition rule for Adv

, used in Theorem 1.5,

3
allows for computing Adv

for a read-once formula by multiplying the bounds computed for

constant-size gates. It may be that the very simple form of our algorithm will allow for further
progress in the understanding and development of quantum query algorithms.
1.1 Open problems
An appealing open problem in quantum computing is to show a tighter relationship between classical
and quantum query complexities for total functionsthe largest known gap is quadratic but the
best upper bound is D(f) = O(Q(f)
6
) [BBC
+
01]. Speculatively, the strong composition properties
of quantum algorithms for total boolean functions may be a tool for approaching this problem. It
also remains interesting to study non-boolean functions, their composition and the necessity of the
log [C[ and log [E[ factors in Eq. (1.5). Theorem 1.5, the two-reections form of the algorithm, and
the elimination of the log Adv

(f) factor suggest that it may be possible to adapt the algorithm to

evaluate any boolean function f with a bounded-error input oracle with the same asymptotic number
(Adv

(f)) of quantum queries, following [HMW03] for the OR function. Classically, in the noisy
decision-tree model, an extra logarithmic factor for error reduction is sometimes required [FRPU94],
but this factor is not known to be needed for any quantum query algorithm [BNRW05].
1.2 Denitions
For a natural number n N, let [n] = 1, 2, . . . , n. For a bit b 0, 1, let

b = 1 b. For a
nite set X, let C
X
be the Hilbert space C
|X|
with orthonormal basis [x : x X. For vector
spaces V and W over C, let /(V, W) denote the set of all linear transformations from V into W,
and let /(V ) = /(V, V ). |A| is the spectral norm of an operator A.
2 The algorithms
For a boolean function, taking the dual of the general adversary bound SDP in Denition 1.1 gives:
Lemma 2.1 ([Rei09a, Theorem 6.2]). Let f : T 0, 1, with T 0, 1
n
. For b 0, 1, let
F
b
= x T : f(x) = b. Then
Adv

(f) = min
mN,[v
xj
C
m
: xT,j[n] :
(x,y)F
0
F
1
,

j[n]:x
j
=y
j
v
xj
[v
yj
=1
max
xT

j[n]
|[v
xj
|
2
.
(2.1)
Based on a feasible solution to this SDP with objective value W( 1), we will give three algorithms
for evaluating f, each with query complexity O(W). (A feasible solution corresponds to a span
program in canonical form, and its value equals the span program witness size [KW93, Rei09a].)
Let I = [n] 0, 1 [m]. Let [t C
F
0
and A /(C
F
0
, C
I
) be given by
[t =
1
3

xF
0
[x
A =

xF
0
,j[n]
[xj, x
j
[ v
xj
[ .
(2.2)
4
Let G be the weighted bipartite graph with biadjacency matrix B
G
/(C
{0}
C
I
, C
F
0
):
B
G
=
_
[t A
_
. (2.3)
That is, G has a vertex for each row or column of B
G
; its vertex set is the disjoint union F
0
0I.
Edges from F
0
to 0 I are weighted by the matrix entries. The weighted adjacency matrix of G is
A
G
=
_
0 B
G
B

G
0
_
. (2.4)
Let /(C
F
0
{0}I
) be the orthogonal projection onto the span of all eigenvalue-zero eigen-
vectors of A
G
. For an input x T, let
x
/(C
F
0
{0}I
) be the projection

x
= 1

j[n],k[m]
[j, x
j
, kj, x
j
, k[ . (2.5)
That is,
x
is a diagonal matrix that projects onto all vertices except those associated to the input
bit complements x
j
. Finally, let
U
x
= (2
x
1)(21) . (2.6)
U
x
consists of the alternating reections 21 and 2
x
1. The rst reection does not depend
on the input x. The second reection can be implemented using a single call to the input oracle O
x
.
We present three related algorithms, each slightly simpler than the one before:
Algorithm 1:
1. Prepare the initial state [0 C
F
0
{0}I
.
2. Run phase estimation on U
x
, with precision
p
=
1
100W
and error rate
e
=
1
10
.
3. Output 1 if the measured phase is zero. Otherwise output 0.
Algorithm 2:
1. Prepare the initial state
1

2
([0 +[1) [0 C
2
C
F
0
{0}I
.
2. Pick T [,100W|] uniformly at random. Apply the controlled unitary [00[
1 +[11[ U
T
x
.
3. Measure the rst register in the basis
1

2
([0[1). Output 1 if the measurement
result is
1

2
([0 +[1), and output 0 otherwise.
Algorithm 3:
1. Prepare the initial state [0 C
F
0
{0}I
.
2. Pick T [,10
5
W|] uniformly at random. Apply U
T
x
.
3. Measure the vertex. Output 1 if the measurement result is [0, and output 0
otherwise.
5
Phase estimation on a unitary V with precision
p
and error rate
e
can be implemented using
O
_
1
p
log
1
e
_
controlled applications of V [NWZ09], so the rst algorithm has O(W) query complexity.
The second algorithm essentially applies a simplied version of phase estimation. Intuitively, it
works because it suces to distinguish zero from nonzero phase. The third algorithm does away with
any phase estimation. Intuitively, this is possible because U
x
is the product of two reections, so its
spectrum is symmetrical. The second and third algorithms clearly have O(W) query complexity.
The factor of 10
5
in the third algorithms query complexity can be reduced by three orders of
magnitude by adjusting downward the scaling factor for [t in Eq. (2.2).
The time, or number of elementary operations, required to implement the reection 21 is
unclear. In practice it may still be preferable to use the potentially less query-ecient quantum
walk algorithm from [Rei09a], as done for evaluating formulas in [Rei09c, R

S08, Rei09d, ACR

+
10].
In the following section, we will show that all three algorithms correctly evaluate f(x), with
constant gaps between the soundness error and completeness parameters.
3 Analysis of the algorithms
To analyze the above algorithms, we shall study the spectrum of the unitary U
x
= (2
x
1)(21).
For this purpose, it will be useful to introduce two new graphs, following [R

S08, Rei09a].
Let (x) =

j[n]
[jj[ [ x
j
x
j
[ 1
C
[m] /(C
I
), and let G(x) and G

(x) be the weighted

bipartite graphs with biadjacency matrices
B
G(x)
=
_
[t A
0 (x)
_
and B
G

(x)
=
_
A
(x)
_
. (3.1)
Based on the constraints of the SDP in Lemma 2.1, we can immediately construct eigenvalue-zero
eigenvectors for G(x) or G

(x), depending on whether f(x) = 1 or f(x) = 0:

Lemma 3.1. If f(x) = 1, let [ = 3

W[0 +

j[n]
[j, x
j
[v
xj
C
{0}I
. Then B
G(x)
[ = 0
and [0[[
2
/|[|
2
9/10.
If f(x) = 0, let [ = [x +

j[n]
[j, x
j
[v
xj
C
F
0
I
. Then B

(x)
[ = 0 and
[t[[
2
/|[|
2
1/(9W(W + 1)).
Let us recall from [Rei09a]:
Theorem 3.2 ([Rei09a, Theorem 8.7]). Let G

be a weighted bipartite graph with biadjacency matrix

B
G
/(C
U
, C
T
). Assume that > 0 and [t, [ C
T
satisfy B

[ = 0 and [t[[
2
|[|
2
.
Let G be the same as G

except with a new vertex, 0, added to the U side, with outgoing edges
weighted by the entries of [t. That is, the biadjacency matrix of G is
B
G
=
_
[t B
G

_
0 U
T (3.2)
Let [ be a complete set of orthonormal eigenvectors of the weighted adjacency matrix A
G
, with
corresponding eigenvalues (). Then for all 0, the squared length of the projection of [0 onto
the span of the eigenvectors with [()[ satises

: |()|
[[0[
2
8
2
/ . (3.3)
6
Substituting Lemma 3.1 into Theorem 3.2, we thus obtain the key statement:
Lemma 3.3. If f(x) = 1, then A
G(x)
has an eigenvalue-zero eigenvector [, supported on the
column vertices, with
[0[[
2
|[|
2

9
10
. (3.4)
If f(x) = 0, let [ be a complete set of orthonormal eigenvectors of A
G(x)
with corresponding
eigenvalues (). Then for any c 0,

:|()|c/W
[[0[
2
72
_
1 +
1
W
_
c
2
. (3.5)
By choosing c a small positive constant, Eq. (3.5) gives an O(1/W) eective spectral gap for
eigenvectors of A
G(x)
supported on [0; it says that [0 has small squared overlap on the subspace
of O(1/W)-eigenvalue eigenvectors.
So far, we have merely repeated arguments from [Rei09a]. The main step in the analysis of our
new algorithms is to translate Lemma 3.3 into analogous statements for U
x
:
Lemma 3.4. If f(x) = 1, then U
x
has an eigenvalue-one eigenvector [ with
[0[[
2
|[|
2

9
10
. (3.6)
If f(x) = 0, let [ be a complete set of orthonormal eigenvectors of U
x
with corresponding
eigenvalues e
i()
, () (, ]. Then for any 0,

:|()|
[[0[
2

_
2

6W +

2
_
2
. (3.7)
Assuming Lemma 3.4, the algorithms from Section 2 are both complete and sound. If f(x) = 1,
then the rst, phase-estimation-based algorithm outputs 1 with probability at least 9/10
e
= 4/5.
If f(x) = 0, then setting =
p
=
1
100W
, the algorithm outputs 1 with probability at most

e
+ (2

6W +

2
)
2
< 2/5. The probability the second algorithm outputs 1 is the expectation
versus T of
1
4
|(1 +U
T
x
)[0|
2
. If f(x) = 1, this is at least 9/10 for all T. If f(x) = 0, let = ,100W|
and simplify
E
T
R
[]
_
1
4
|(1 +U
T
x
)[0|
2
_
= E
T
R
[]
_
1
4

[1 +e
i()T
[
2
[0[[
2
_
=
1
4

[0[[
2
_
2 +
1

_
e
i()(+1)
e
i()
e
i()
1
1
__
.
(3.8)
Setting =
1
50W
and = (2

6W +

2
)
2
, we see that the algorithm outputs 1 with probability at
most + (1 )
_
1
2
+ 1/(4 sin

2
)
_
< 88% for W 1. As its analysis requires more care, we defer
consideration of the third algorithm to the end of this section.
For the proof of Lemma 3.4 we will use the following characterization of the eigen-decomposition
of the product of reections, essentially due to Jordan [Jor75]. Its use is common in quantum
computation, e.g., [NWZ09, Sze04, MW05].
7
Lemma 3.5. Given two projections and , the Hilbert space can be decomposed into orthogonal
one- and two-dimensional subspaces invariant under and . On the one-dimensional invariant
subspaces, (21)(21) acts as either +1 or 1. Each two-dimensional subspace is spanned by
an eigenvalue- eigenvector [v of , with (0, 1), and [v

= (1 )[v/|(1 )[v|.
Letting = 2 arccos

(0, ), so [v/|[v| = cos

2
[v + sin

2
[v

, the eigenvectors and

corresponding eigenvalues of (21)(21) on this subspace are, respectively,
[v i[v

2
and e
i
. (3.9)
Proof of Lemma 3.4. Notice from Eqs. (2.3) and (3.1) that G is naturally a subgraph of G(x).
Since A
G
= 0 by denition of , A
G(x)
= T(1
x
), where T is a permutation matrix.
First consider the case f(x) = 1. Let [ be the restriction of [ from Eq. (3.4) to the
vertices of G. Since [ has no support on the extra vertices of G(x), |[| = |[| and [ is an
eigenvalue-zero eigenvector of A
G
; [ = [. Also
x
[ = [, so indeed U
x
[ = [.
Next consider the case f(x) = 0. Let
[ =

:|()|
[[0 (3.10)
be the projection of [0 onto the space of eigenvectors with phase at most in magnitude. Our aim
is to upper bound |[|
2
= 0[ = [0[

[
2
, where [

= [/|[|. Notice that [

is supported only
on eigenvectors [ with () ,= 0, i.e., on the two-dimensional invariant subspaces of
x
and .
Indeed, if U
x
[ = [, then either [ =
x
[ = [ or [ = (1
x
)[ = (1 )[. The rst
possibility implies A
G(x)
[ = 0, so by Eq. (3.5) with c = 0, 0[ = 0. In the second possibility,
also 0[ = 0[
x
[ = 0 since [0 =
x
[0.
We can split 0[

as
0[

= 0[[

+0[
x
(1 )[

[0[[

[ +[0[
x
(1 )[

[
[0[[

[ +|
x
(1 )[

| .
(3.11)
Start by bounding the second term, |
x
(1 )[

|. Intuitively, this term is small because [

is supported only on two-dimensional invariant subspaces where and

x
are close. Indeed, let
[ = (21)[, an eigenvector of A
G
with phase () = (). Expanding [

[,
|
x
(1 )[

|
2
= |

x
(1 )c

[|
2
=

:()>0
|
x
(1 )(c

[ +c

[)|
2
=

:()>0
sin
2
()
2
|(1 )(c

[ +c

[)|
2

2
_
2
|(1 )[

|
2
. (3.12)
It remains to bound [0[[

[ = [0[w[|[

|, where [w = [

/|[

| is an eigenvalue-zero
eigenvector of A
G
. Intuitively, if [0[w[ = [0[
x
[w[ is large, then since A
G
and A
G(x)
are the
8
same on
x
, |A
G(x)
[w| = |T(1
x
)[w| will be small. This in turn will imply that [w has large
support on the small-eigenvalue subspace of A
G(x)
, contradicting Eq. (3.5).
Beginning the formal argument, we have
|A
G(x)
[

|
2
= |(1
x
)[

|
2
=

:()>0
|(1
x
)(c

[ +c

[)|
2
=

:()>0
sin
2
()
2
|(c

[ +c

[)|
2

2
_
2
|[

|
2
.
(3.13)
Hence |A
G(x)
[w| /2.
Now split [w = [w
small
+[w
big
, where for a certain d > 0 to be determined,
[w
small
=

:|()|d/2
[[w
[w
big
=

:|()|>d/2
[[w .
(3.14)
Then
[0[[

[ [0[w[ [0[w
small
[ +[0[w
big
[ . (3.15)
From Eq. (3.5) with c = dW/2, [0[w
small
[
_
72(1 + 1/W)c|[w
small
| 6dW.
Since A
G(x)
[w =

()[[w, we have
_

2
_
2
|A
G(x)
[w|
2
= |A
G(x)
[w
small
|
2
+|A
G(x)
[w
big
|
2
d
2
_

2
_
2
|[w
big
|
2
.
(3.16)
Hence |[w
big
| 1/d.
Combining our calculations gives

:|()|
[[0[
2
= 0[

[0[[

[ +|
x
(1 )[

| 6dW +
1
d
+

2
.
(3.17)
The right-hand side is 2

6W + /2, as claimed, for d = 1/

6W.
Having proved Lemma 3.4, we return to the correctness proof for the third algorithm.
Proposition 3.6. If f(x) = 1, then the third algorithm outputs 1 with probability at least 64%.
If f(x) = 0, then the third algorithm outputs 1 with probability at most 61%.
9
Proof. Letting = ,10
5
W|, the third algorithm outputs 1 with probability
p
1
:= E
T
R
[]
_
[0[U
T
x
[0[
2

= E
T
R
[]

e
i()T
[[0[
2

2
. (3.18)
If f(x) = 1, then a crude bound puts p
1
at least (9/10 1/10)
2
= 64%.
Assume f(x) = 0. Recall the notation that for an eigenvector [ with [()[ (0, ), [ =
(21)[ denotes the corresponding eigenvector with eigenvalue phase () = (). The key
observation for this proof is that
0[ = e
i()
0[ . (3.19)
This equal splitting of [0[[ and [0[[ will allow us to bound p
1
close to 1/2 instead of the trivial
bound p
1
1. The intuition is that after applying U
x
a suitable number of times, eigenvectors [
and [ will accumulate roughly opposite phases, so their overlaps with [0 will roughly cancel out.
For this argument to work, though, the eigenvalue phase () should be bounded away from zero
and from . Therefore dene the projections

:|()|
[[

:|()|>
[[
= 1

,
(3.20)
where and , 0 < < < , will be determined below. Lemma 3.4 immediately gives the
bound |

[0| 2

6W +

2
. We can also place a bound on |

[0|, using
2(1)[0 = (U

x
1)[0 =

(e
i()
1)[[0 . (3.21)
Expanding the squared norm of both sides gives
|(U

x
1)[0|
2
= 4

sin
2
()
2
[[0[
2
|

[0|
2
4 sin
2

2
(3.22)
and
|(U

x
1)[0|
2
= 4(1 |[0|
2
) 2/5 . (3.23)
In the second step we have used that |[0|
2
9/10; provided that f is not the constant zero
function, A
G
must have an eigenvalue-zero eigenvector with large overlap on [0. Combining
Eqs. (3.22) and (3.23) gives
|

[0|
2

1
10 sin
2
2
. (3.24)
Returning to Eq. (3.18), we have
p
1
E
T
R
[]
_
|

[0|
2
+|

[0|
2
+

:()(,]
[[0[
2
_
e
i()T
+e
i()T
_

_
2
(|

[0|
2
+|

[0|
2
)(2 |

[0|
2
|

[0|
2
)
+ E
T
R
[]
_

:()(,]
[[0[
2
_
e
i()T
+e
i()T
_
_
2
.
(3.25)
10
The algorithm chooses T at random to allow bounding the last term. Expanding this term gives
E
T
R
[]

:(),(

)(,]
[[0[
2
[

[0[
2
_
e
i()T
+e
i()T
__
e
i(

)T
+e
i(

)T
_
= E

(,]
[[0[
2
[

[0[
2
_
(e
i(+

)T
+e
i(+

)T
) + (e
i(

)T
+e
i(

)T
)
_

1
2
|[0|
4
+ E

(,]
[[0[
2
[

[0[
2
_
e
i(+

)T
+e
i(+

)T
_
=
1
2
|[0|
4
+
1

(,]
[[0[
2
[

[0[
2
_
e
i(+

)(+1)
e
i(+

)
e
i(+

)
1
1
_

1
2
_
1 +
1/
min
,

(,]
[e
i(+

)
1[
_
|[0|
4

1
2
_
1 +
1
2 minsin , sin
_
|[0|
4
.
(3.26)
Here for brevity we have written and

for () and (

), respectively. In the second and fourth

steps, we have used

(,]
[[0[
2
=
1
2
|[0|
2
. In the last step, we have used [e
i(+

)
1[ =
2 sin
+

2
2 minsin , sin . Substituting the result back into Eq. (3.25) gives
p
1
1
1
2
_
1
1
2 minsin , sin
_
|[0|
4
1
1
2
_
1
1
2 minsin , sin
_
max
_
1
1
10 sin
2
2

_
2

6W +

2
_
2
, 0
_
2
.
(3.27)
Setting = and = 1/(2000W), for W 1 a calculation yields p
1
61%.
Acknowledgements
I would like to thank Troy Lee for his help in formulating Theorem 1.2. I also thank Sergio Boixo,
Stephen Jordan, Julia Kempe and Rajat Mittal for helpful comments, and the Institute for Quantum
Information for hospitality. Research supported by NSERC, ARO and MITACS.
References
[AA09] Scott Aaronson and Andris Ambainis. The need for structure in quantum speedups.
2009, arXiv:0911.0996 [quant-ph].
[Aar09] Scott Aaronson. BQP and the polynomial hierarchy. 2009, arXiv:0910.4698
[quant-ph].
[ACR
+
10] Andris Ambainis, Andrew M. Childs, Ben W. Reichardt, Robert

Spalek, and Shengyu
Zhang. Any AND-OR formula of size N can be evaluated in time N
1/2+o(1)
on a quantum
computer. SIAM J. Comput., 39(6):25132530, 2010. Earlier version in FOCS07.
[Amb02] Andris Ambainis. Quantum lower bounds by quantum arguments. J. Comput. Syst.
Sci., 64:750767, 2002, arXiv:quant-ph/0002066. Earlier version in STOC00.
11
[Amb07] Andris Ambainis. A nearly optimal discrete query quantum algorithm for evaluating
NAND formulas. 2007, arXiv:0704.3628 [quant-ph].
[AS04] Scott Aaronson and Yaoyun Shi. Quantum lower bounds for the collision and the
element distinctness problem. J. ACM, 51(4):595605, 2004.
[BBC
+
01] Robert Beals, Harry Buhrman, Richard Cleve, Michele Mosca, and Ronald de
Wolf. Quantum lower bounds by polynomials. J. ACM, 48(4):778797, 2001,
arXiv:quant-ph/9802049. Earlier version in FOCS98.
[BHT98] Gilles Brassard, Peter Hyer, and Alain Tapp. Quantum algorithm for the col-
lision problem. In Proc. 3rd LATIN, LNCS vol. 1380, pages 163169, 1998,
arXiv:quant-ph/9705002.
[BNRW05] Harry Buhrman, Ilan Newman, Hein R ohrig, and Ronald de Wolf. Robust polynomials
and quantum algorithms. In Proc. 22nd STACS, LNCS vol. 3404, pages 593604, 2005,
arXiv:quant-ph/0309220.
[BSS03] Howard Barnum, Michael Saks, and Mario Szegedy. Quantum query complexity and
semidenite programming. In Proc. 18th IEEE Complexity, pages 179193, 2003.
[BV97] Ethan Bernstein and Umesh Vazirani. Quantum complexity theory. SIAM J. Comput.,
26(5):14111473, 1997. Earlier version in STOC93.
[BW02] Harry Buhrman and Ronald de Wolf. Complexity measures and decision tree complexity:
A survey. Theoretical Computer Science, 288(1):2143, 2002.
[CGM
+
09] Richard Cleve, Daniel Gottesman, Michele Mosca, Rolando Somma, and David L. Yonge-
Mallo. Ecient discrete-time simulations of continuous-time quantum query algorithms.
In Proc. 41st ACM STOC, pages 409416, 2009, arXiv:0811.4428 [quant-ph].
[CL08] Andrew M. Childs and Troy Lee. Optimal quantum adversary lower bounds for ordered
search. In Proc. 35th ICALP, LNCS vol. 5125, pages 869880, 2008, arXiv:0708.3396
[quant-ph].
[FRPU94] Uriel Feige, Prabhakar Raghavan, David Peleg, and Eli Upfal. Computing with noisy
information. SIAM J. Comput., 23(5):10011018, 1994. Earlier version in STOC90.
[Gro96] Lov K. Grover. A fast quantum mechanical algorithm for database search. In Proc.
28th ACM STOC, pages 212219, 1996, arXiv:quant-ph/9605043.
[HL

S07] Peter Hyer, Troy Lee, and Robert

Spalek. Negative weights make adversaries stronger.
In Proc. 39th ACM STOC, pages 526535, 2007, arXiv:quant-ph/0611054.
[HMW03] Peter Hyer, Michele Mosca, and Ronald de Wolf. Quantum search on bounded-error
inputs. In Proc. 30th ICALP, pages 291299, 2003, arXiv:quant-ph/0304052. LNCS
2719.
[H

S05] Peter Hyer and Robert

Spalek. Lower bounds on quantum query complexity. EATCS
Bulletin, 87:78103, October 2005, arXiv:quant-ph/0509153.
12
[Jor75] Camille Jordan. Essai sur la geometrie ` a n dimensions. Bulletin de la S. M. F., 3:103174,
1875.
[KW93] Mauricio Karchmer and Avi Wigderson. On span programs. In Proc. 8th IEEE Symp.
Structure in Complexity Theory, pages 102111, 1993.
[Lee09] Troy Lee. Composition upper bound for half-Boolean functions. unpublished, 2009.
[MW05] Chris Marriott and John Watrous. Quantum Arthur-Merlin games. Computational
Complexity, 14(2):122152, 2005, arXiv:cs/0506068 [cs.CC].
[NWZ09] Daniel Nagaj, Pawel Wocjan, and Yong Zhang. Fast amplication of QMA. Quantum
Inf. Comput., 9:10531068, 2009, arXiv:0904.1549 [quant-ph].
[Rei09a] Ben W. Reichardt. Span programs and quantum query complexity: The general
adversary bound is nearly tight for every boolean function. 2009, arXiv:0904.2759.
[Rei09b] Ben W. Reichardt. Span programs and quantum query complexity: The general
adversary bound is nearly tight for every boolean function. In Proc. 50th IEEE FOCS,
pages 544551, 2009.
[Rei09c] Ben W. Reichardt. Span-program-based quantum algorithm for evaluating unbalanced
formulas. 2009, arXiv:0907.1622 [quant-ph].
[Rei09d] Ben W. Reichardt. Faster quantum algorithm for evaluating game trees. 2009,
arXiv:0907.1623 [quant-ph].
[Rei10] Ben W. Reichardt. Least span program witness size equals the general adversary
lower bound on quantum query complexity. Technical Report TR10-075, Electronic
Colloquium on Computational Complexity, http://eccc.hpi-web.de, 2010.
[R

S08] Ben W. Reichardt and Robert

Spalek. Span-program-based quantum algorithm for
evaluating formulas. In Proc. 40th ACM STOC, pages 103112, 2008, arXiv:0710.2630
[quant-ph].
[Sho97] Peter W. Shor. Polynomial-time algorithms for prime factorization and discrete
logarithms on a quantum computer. SIAM J. Comput., 26(5):14841509, 1997,
arXiv:quant-ph/0508027. Earlier version in FOCS94.
[Sim97] Daniel R. Simon. On the power of quantum computation. SIAM J. Computing,
26(5):14741483, 1997. Earlier version in FOCS94.
[

SS06] Robert

Spalek and Mario Szegedy. All quantum adversary methods are equivalent.
Theory of Computing, 2(1):118, 2006, arXiv:quant-ph/0409116. Earlier version in
ICALP05.
[Sze04] Mario Szegedy. Quantum speed-up of Markov chain based algorithms. In Proc. 45th
IEEE FOCS, pages 3241, 2004.
Institute for Quantum Computing, University of Waterloo
E-mail address: [email protected]
13